What's new in Codron

PREVIOUS

1. v0.25.1Oct 1, 2026A seat comes before the invite, and seats change from your account page

   • · An open invite holds a Team seat until it is accepted, revoked or expires. When every seat is taken, the Team members panel offers Add a seat instead of a new invite link. A teammate already on another of your projects already holds a seat, so you can still invite them by email.
   • · Seats change on your codron.app account page. Added seats are ready at once, and the rest of the billing period is charged then. Fewer seats take effect at the next billing date, with no refund, and only down to the people on your team and open invites, so remove someone first to go lower. Until that date the seats being given up can't be filled.
   • · If a team ever has more members than seats, its team workspaces pause, with members and discussions kept, and reopen as soon as seats are added or people are removed. Settings → Account and the Team members panel say how many.
   • · A refused invite says why in the app instead of showing a generic error.
2. v0.25.0Oct 1, 2026Personal use is free with no limits, the app speaks Korean, and Claude sign-in stays with Claude Code

   • · Personal use is free. The one-project limit is gone, and so are the Team locks on Vercel, Google Cloud, ngrok and custom integrations. The paid plan is now Team, billed per seat, for sharing projects with other people. Accepting an invite on a team with no free seat says so and asks the owner to add one.
   • · The app speaks Korean. Settings has a Language choice: System, English or 한국어, where System follows your computer. Onboarding, Overview, Projects, a project's results, decisions and request queue, New project and Settings are translated, and a language picked in the Preferences window reaches the main window when you return to it.
   • · Claude sign-in runs claude auth login, and Codron no longer keeps its own copy of your Claude token. If you only ever signed in to Claude through Codron, sign in once more from Settings → Engines after this update. A run that finds Claude signed out now asks you to sign in instead of leaving the project stuck, and the sign-in page opens once, from Claude Code itself. Signing in or out here does the same for Claude Code everywhere on that computer, the terminal included.
   • · The architecture map is an interactive page with its own section in the Docs tab. Search for a part, trace what calls what, play the guided chapters and export PNG or SVG. Draw.io no longer needs to be installed, and a draw pressed right after Codron starts is no longer lost.
   • · Runs show the sub-agents they start, each in its own row with its status, steps and result, and the running session card counts the ones still working. With smart routing on, those helpers take the lighter tier while the reviewer stays on the run's model.
   • · Smart routing works on Windows again. Since September 12 every Claude classification there failed, so each run fell back to the primary engine.
   • · The Source tab reads the checkout on this computer, so a project last run on another machine no longer shows as not initialized.
   • · On macOS this is the first build since 0.24.0, so it also brings the 0.24.1 changes below.
3. v0.24.1Sep 23, 2026Source changes and history fold and scroll on their own, the machine cell names the other machines, and the Docs tab drops the agent's notes

   • · Changes and history in the Source view fold like the source control panes in VS Code. Each has its own header and its own scroll, so a long list of changed files no longer pushes the history off screen. A folded pane gives its space to the open one, and what you fold stays folded across projects. The changes header counts files, the history header counts unpushed commits, and a failed pull or push shows right under the buttons.
   • · History no longer stops at 50 commits without saying so. Load more adds 50 at a time up to the latest 200, and the list says when it has reached them.
   • · The projects list names the other machines that have a project. With a machine filter on, every row used to repeat that machine's name with +1; the cell now reads also and the other machine's name, or stays empty when the project lives only there. Under All machines it lists the machine names, the one that ran last first.
   • · The Docs tab no longer shows the open frontier. Those items are the agent's own notes for its next cycle, so a finished project kept reading as open work you had nothing to do about. The agent still keeps and uses them.
4. v0.24.0Sep 23, 2026Models follow the installed CLI, Opus 5.5 and GPT-6 Sol by default, and a spent model limit says so

   • · The model and effort pickers follow the CLI you have installed. Claude Code and Codex both keep their model catalog on disk, and Codron now reads it, so a model the vendor ships appears in Settings, the project override, New project and the chat picker once the CLI has run, with its effort tiers and the CLI version it needs. Nothing in Codron has to update first.
   • · New defaults: Claude Opus 5.5 at xhigh and GPT-6 Sol at high, each newer and cheaper than the model it replaces. Settings and projects saved on opus-5 or gpt-5.6-sol move forward on their own; a model you picked yourself, such as Fable or Astra, stays. Opus 5.5 needs Claude Code 2.1.280, and an older CLI runs Opus 5 for that run and says to update.
   • · Smart routing is on by default. Simple requests go to a lighter tier and complex work stays on your primary engine; on Codex the tiers are now GPT-6 Luna and GPT-6 Sol. A project that had routing off before this release is switched on, and turning it off again in Settings now saves that choice.
   • · Independent AI review is gone. In every recorded run the primary engine never asked for it, and both engines already check their own work. Saved settings that still carry a reviewer keep working.
   • · A spent model limit says so. Claude reports the Fable weekly limit only as a meter held at 99%, which the usage panel read as still serving while every Fable request was refused. That meter now reads as reached, with a Fable limit reached badge, and Codex per-model limits get the same treatment.
   • · A run that hits one model's own limit moves to a model that still serves until the reset, and says so in the activity log. Weekly reset times are read correctly now, so a project waiting on an account limit sleeps until the real reset instead of retrying every hour.
5. v0.23.13Sep 21, 2026Rescue agent output is labeled everywhere, usage says when a provider has actually stopped, and the dashboard loads without rebuilding itself

   • · Every line a rescue agent writes is marked [rescue] and shown in amber, now in the dev server and mobile build logs as well as the tunnel and engine install logs. A rescue agent edits your repository unattended, so what it says should not read as ordinary tool output. Codron's own announcements, the warning, the retry and the give-up, stay unmarked.
   • · A usage window says when a provider has actually stopped serving. Both CLIs report this and the panel was throwing it away, so a week that is full but still answering turned red at 95 percent, while a workspace that had run out of credits looked like nothing more than a full bar. The meter now takes its color and its limit reached wording from the provider, not from the percentage.
   • · The dashboard stops rebuilding parts of itself as it loads. A donut tooltip, the sidebar project age and five other relative times came out one way on the server and another in the browser, which threw the first paint away. Relative ages now wait for the browser, so the two agree. A settings row also drew its buttons across its own hint in a narrow panel; the row wraps now and the buttons take their own line.
6. v0.23.12Sep 17, 2026A refused GitLab token is reported as one, and a slow status check no longer reads as a sign-out

   • · A GitLab token the server refuses is reported as a failure. glab keeps such a token and exits cleanly with only a warning, so Codron used to end the sign-in with done while the row stayed on needs sign-in. It now reads the warning, says under the token field what to check (copied whole from the one-time reveal, not expired or revoked, api scope) and that the row stays on needs sign-in until a working token replaces the stored one.
   • · A slow status check no longer reads as a sign-out. Some CLIs take two to three seconds just to start on Windows, and Settings checks nine at once; at the old three second limit the Vercel row kept flipping to needs sign-in on a signed-in account. Every check now gets ten seconds, and a check that does not answer keeps the row on its last known state. A real sign-out answers at once and still shows.
7. v0.23.11Sep 17, 2026Settings opens on Machines, the GitLab card explains a blocked browser sign-in, and cancel acts on the click

   • · Settings opens on Machines, which now leads the list: which computer this is and its Codron version, then Engines, Integrations and Variables on this machine, then the Workspace where it puts projects. Links that open a specific tab keep working.
   • · The GitLab server card offers the browser sign-in only where it can start. On a company server Codron asks glab whether an OAuth client ID is stored for it; without one the card says so in plain text and points to the token above, instead of a link that fails into the log. If an attempt is still refused, the explanation appears in the card and no exit code error is added. Log panels wrap long lines so guidance reads in full.
   • · cancel on a CLI sign-in acts on the click. The button appears only while the CLI runs, leaves the moment you click it, and the row is freed as soon as the CLI is gone; the status check afterwards runs in the background. glab's status check also gets more time to start, so a signed-in account no longer shows needs sign-in on a slow machine.
8. v0.23.10Sep 17, 2026GitLab clone protocol applies on the click, and one sign-in per server

   • · The Clone with choice in the GitLab server card applies on the click. Each click used to wait on a glab status probe, and glab takes two to three seconds to start on some Windows computers; a second click inside that wait was overwritten when the first reply landed late, so the radio jumped back and forth. Codron now saves the choice and returns at once, updates glab in the background with quick clicks collapsed to the newest one, and reverts the radio only when the save itself is refused.
   • · One sign-in per GitLab server. On a company server the card leads with Sign in with token, the GitLab CLI row hides its browser button, and the browser flow stays as a link with its condition spelled out: glab only carries an OAuth client ID for gitlab.com, so that flow needs an application registered on your server. On gitlab.com the row keeps its sign in button and the token form sits behind a link.
9. v0.23.9Sep 16, 2026Stop is instant, self-managed GitLab sign-in, and parked projects recover a purged session

   • · Stop on a dev server ends the process on the click. The desktop app now sends the stop straight down to its own agent instead of a round trip out to the cloud and back, so the port frees on the same tick. The card and the preview both drop to stopped right away, and the preview refreshes itself when its tab regains focus instead of leaving a dead page on screen.
   • · A project parked for weeks runs again instead of failing in seconds. Claude Code removes old sessions on its own schedule, and a lane still holding a removed session id was rejected at resume. The agent retries once in the same cycle with a fresh session and the full prompt, returns a preview snap that produced no work to the queue rather than marking it failed, and keeps requests written into the project file open for that retry.
   • · Sign into a self-managed GitLab with a personal access token. The device flow only carries a client id for gitlab.com, so a company server stopped with a configuration error before it ever printed a URL. The GitLab server card in Settings takes a token, hands it to glab over standard input so it never reaches the log, and names that refusal in the row when it happens.
   • · Choose SSH or HTTPS for GitLab clones, per computer. glab cloned with whatever protocol its host configuration held, which left private HTTPS clones to Git's credential store. The choice applies at sign-in, and switches the host setting when you are already signed in.
10. v0.23.8Sep 15, 2026CLI installs show up without a restart, sign-ins finish on screen, Codex runs record their commands

    • · Settings sees a CLI installed since Codron started. Windows installers add their folder to the system PATH, which the running app never re-read, so ngrok and other winget installs stayed on "not installed" while the rescue agent kept reporting success. The app now re-reads the PATH before every check and after every install, treats winget's already-installed result as installed, and adopts the folder the rescue agent reports.
    • · Installs run one at a time. Click every pending update at once; a waiting row says what it is queued behind in its log, so two npm or winget installs never collide and a claude update never interrupts a rescue agent running on it. Rows still installing keep their status and update badge while another finishes.
    • · Sign-ins finish on screen. Settings and onboarding no longer wait for the CLI to exit: the spinner ends when the token appears or the CLI reports signed in, a cancel button stops a login whose browser step never came back, and a login completed through the CLI's own store replaces a stale saved token. Covers claude, codex, gh, glab, gcloud, vercel, supabase and library integrations.
    • · Codex runs record their shell commands in run evidence. When Git evidence is missing, the panel says why (no repository, Git not installed, timeout) and what to do. A collapsed run card shows only the result, with the requests it picked up under the report.
    • · Usage chart: the hourly tooltip stays inside the plot, and a tiny engine share next to a large one stays visible.
    • · macOS: the window controls no longer cover the sidebar logo.